The Existing DCT-Based JPEG Standard 1
What Is JPEG? The JPEG (Joint Photographic Experts Group) committee, formed in 1986, has been chartered with the Digital compression and coding of continuous-tone still images Joint between ISO and ITU-T Has developed standards for the lossy, lossless, and nearly lossless of still images in the past decade Website: www.jpeg.org The JPEG committee has published the following standards: ISO/IEC 10918-1 ITU-T Rec. T.81 : Requirements and guidelines ISO/IEC 10918-2 ITU-T Rec. T.83 : Compliance testing ISO/IEC 10918-3 ITU-T Rec. T.84: Extensions ISO/IEC 10918-4 ITU-T Rec. T.86: Registration of JPEG Parameters, Profiles, Tags, Color Spaces, APPn Markers Compression Types, and Registration Authorities (REGAUT) 2
The Existing JPEG Standard Toolkit The existing JPEG standard concerns with the compression of continuous-tone, still-frame, monochrome and color images. It provides a "toolkit" of compression techniques from which applications can select the elements that satisfy their particular requirements. Baseline system: A simple and efficient DCT-based algorithm that uses Huffman coding, operates only in sequential mode, and is restricted to 8 bits/pixel input. Extended system: Enhancements to the baseline to satisfy broader applications. Lossless mode: Based on predictive coding and independent of the DCT that uses either Huffman or arithmetic coding. 3
JPEG Encoder Block Diagram 8x8 Blocks Header Compressed Data FDCT Quantizer Huffman Encoder Quantization Tables Huffman Tables 4
JPEG Decoder Block Diagram Header Compressed Data Reconstructed Image Data Huffman Decoder Dequantizer IDCT Huffman Tables Quantization Tables 5
Image Representation with DCT DCT coefficients can be viewed as weighting functions that, when applied to the 64 cosine basis functions of various spatial frequencies (8 x 8 templates), will reconstruct the original block. = y(0,0) x + y(1,0) x + + y(7,7) x Original image block DC (flat) basis function AC basis functions 6
Original Image JPEG DCT Example 139 144 149 153 155 155 155 155 144 151 153 156 159 156 156 156 150 155 160 163 158 156 156 156 159 161 162 160 160 159 159 159 159 160 161 162 162 155 155 155 161 161 161 161 160 157 157 157 162 162 161 163 162 157 157 157 162 162 161 161 163 158 158 158 8 BPP 64 pixels 512 bits DCT Transformed Image 235.6-1.0-12.1-5.2 2. 1-1. 7-2.7 1.3-22. 6-17.5-6.2-3.2-2.9-0. 1 0.4-1.2-10. 9-9.3-1.6 1.5 0. 2-0. 9-0.6-0.1-7. 1-1.9 0.2 1.5 0. 9-0. 1 0.0 0.3-0. 6-0.8 1.5 1.6-0.1-0. 7 0.6 1.3 1.8-0.2 1.6-0.3-0.8 1.5 1.0-1.0-1. 3-0.4-0.3-1.5-0.5 1.7 1.1-0.8-2. 6 1.6-3.8-1.8 1. 9 1.2-0.6-0.4 Quantized/ Scaled Transformed Data 15 0-1 0 0 0 0 0-2 -1 0 0 0 0 0 0-1 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 7
DCT Coefficient Quantization Each DCT coefficient is uniformly quantized with a quantization step that is taken from a user-defined quantization table (q-table or normalization matrix), characterized by 64, 1-byte elements. The quality and compression ratio of an encoded image can be varied by changing the q-table elements (usually by scaling up or down the values of an initial q-table). The q-table is often designed according to the perceptual importance of the DCT coefficients (e.g., by using the HVS CSF data) under the intended viewing conditions. For the baseline system, in order to meet the needs of the various color components, up to four different quantization tables are allowed. 8
Example of Luminance Quantization Table The JPEG committee has listed the following luminance quantization table as an example in Annex K (informative) of the IS. It is obtained by measuring the DCT coefficient visibility threshold using CCIR-601 images and display, at a distance of six picture-heights away. QL( u, v) = 16 11 10 16 24 40 51 61 12 12 14 19 26 58 60 55 14 13 16 24 40 57 69 56 14 17 22 29 51 87 80 62 18 22 37 56 68 109 103 77 24 35 55 64 81 104 113 92 49 64 78 87 103 121 120 101 72 92 95 98 112 100 103 99 9
JPEG DCT Example After Quantization, the DCT is separated into a DC coefficient and AC coefficients, which are reordered into a 1- D format using a zigzag pattern in order to create long runs of zero-valued coefficients. The DC coefficient is directly correlated to the mean of the 8-by-8 block (upperleft corner). All DC coefficients are combined into a separate bit stream. The AC coefficients are the values of the cosine basis functions (all other values). DC 10
JPEG DCT Example DC AC 15 0-2 -1-1 -1 0 0-1 EOB The DC coefficient is encoded using Huffman encoded 1D- DPCM The AC coefficients are encoded using Huffman coding on magnitude/runlength pairs (magnitude of a nonzero AC coefficient plus runlength of zero-valued AC coefficients that precede it). The end-of-block (EOB) symbol indicates that all remaining coefficients in the zigzag scan are zero. This allows many coefficients to be encoded with only a single symbol. 11
JPEG DCT Example Dequantized DCT Coefficients 240 0-10 0 0 0 0 0-24 -12 0 0 0 0 0 0-14 -13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Reconstructed Image 144 146 149 152 154 156 156 156 148 150 152 154 156 156 156 156 155 156 157 158 158 157 156 155 160 161 161 162 161 159 157 155 163 163 164 163 162 160 158 156 163 163 164 164 162 160 158 157 160 161 162 162 162 161 159 158 158 159 161 161 162 161 159 158 Error Image -5-2 0 1 1-1 -1-1 -4 1 1 2 3 0 0 0-5 -1 3 5 0-1 0 1-1 0 1-2 -1 0 2 4-4 -3-3 -1 0-5 -3-1 -2-2 -3-3 -2-3 -1 0 2 1-1 1 0-4 -2-1 4 3 0 0 1-3 -1 0 RMSE= 2.26 12
The Emerging JPEG2000 Standard 13
JPEG-DCT Pros and Cons Advantages Memory efficient Low complexity Compression efficiency Visual model utilization Robustness Disadvantages Single resolution Single quality No target bit rate No lossless capability No tiling No region of interest Blocking artifacts Poor error resilience 14
JPEG2000 Compression Paradigm Encode Decode Decode Choices Image resolution SNR fidelity Visual fidelity Target filesize Lossless/lossy Region-of-interest Tiles 15
Embedded Bitstream Application Lossless High quality Image 1 Image 2 Image 3 Image 4 16
JPEG2000 Objectives Advanced standardized image coding system to serve applications into the next millenium Address areas where current standards fail to produce the best quality or performance Provide capabilities to markets that currently do not use compression Provide an open system approach to imaging applications 17
JPEG2000: Requirements and Profiles Internet applications (World Wide Web imagery) Progressive in quality and resolution, fast decode Mobile applications Error resilience, low power, progressive decoding Digital photography Low complexity, compression efficiency Hardcopy color facsimile, printing and scanning Compression efficiency, strip or tile processing Digital library/archive applications Metadata, content management Remote sensing Multiple components, fast encoding, region of interest Medical applications Region of interest coding, lossy to lossless 18
JPEG2000 Features Improved compression efficiency (estimated 5-30% depending on the image size and bit rate) Lossy to lossless Multiple resolution Embedded bit stream (progressive decoding) Region of interest coding (ROI) Error resilience Bit stream syntax File format 19
JPEG2000 Compression Standard The standard only specifies a decoder and a bitstream syntax and is issued in several parts: Part I: specifies the minimum compliant decoder (e.g., a decoder that is expected to satisfy 80% of applications); International Standard (IS) is expected 12/2000. Part II: Describes optional, value added extensions; IS is expected in 12/2001. Other parts include: Motion JPEG2000 (Part III); Conformance Testing (Part IV); reference software in JAVA and C (Part V); file format for compound images (Part VI); and Technical Report outlining guidelines for minimum support of Part I (Part VII). 20
JPEG2000 Part I Encoder Original Image Data Discrete Wavelet Transform (DWT) Uniform Quantizer with Deadzone Compressed Image Data Adaptive Binary Arithmetic Coder 21
Benefits of DWT Multiple resolution representation Lossless representation with integer filters Better decorrelation than DCT, resulting in higher compression efficiency Use of visual models DWT provides a frequency band decomposition of the image where each subband can be quantized according to its visual importance (similar to the quantization table specification in JPEG-DCT) 22
JPEG2000 DWT Choices JPEG-2000 Part I only allows successive powers of two splitting of the LL band and the use of two DWT filters: The integer (5,3) filter provides fast implementation (faster than DCT) and lossless capability, but at the expense of some loss in coding efficiency. The Daubechies (9,7) floating-point filter that provides superior coding efficiency. The analysis filters are normalized to a DC gain of one and a Nyquist gain of 2. Part II allows for arbitrary size filters (user-specified in the header), arbitrary wavelet decomposition trees, and different filters in the horizontal vs. vertical directions. 23
The 1-D Two-Band DWT x(n) low-pass h 0 2 low-pass output h 1 2 high-pass high-pass output Analysis filter bank 24
Example of Analysis Filter Bank 1-D signal: 100 100 100 100 200 200 200 200... Low-pass filter h 0 : (-1 2 6 2-1)/8 High-pass filter h 1 : (-1 2-1)/2 Before downsampling: 100 100 87.5 112.5 187.5 212.5 200 200... 0 0 0-50 50 0 0 0 After downsampling: 100 112.5 212.5 200... 0 0 50 0 25
26
27
28
Inverse DWT During the inverse DWT, each subband is interpolated by a factor of two by inserting zeros between samples and then filtering each resulting sequence with the corresponding lowpass, g 0, or high-pass, g 1, synthesis filters. The filtered sequences are added together to form an approximation to the original signal. 0 100 0 112.5 0 212.5 0 200 0 0 0 0 50 0 0 0 100 100 100 100 200 200 200 200... 29
The 1-D Two-Band DWT low-pass low-pass x(n) h 0 2 2 g 0 y(n) h 1 2 2 g 1 high-pass Analysis filter bank high-pass Synthesis filter bank Ideally, it is desired to choose the analysis filter banks (h 0 and h 1 ), and the synthesis filter banks (g 0 and g 1 ), in such a way so as to make the overall distortion zero, i.e., x(n) = y(n). This is called the perfect reconstruction property. 30
JPEG2000 Part I Encoder Original Image Data Discrete Wavelet Transform (DWT) Uniform Quantizer with Deadzone Compressed Image Data Adaptive Binary Arithmetic Coder Quantization step size can vary from one subband to another according to visual models (similar to JPEG Q- table specification). 31
Quantization in Part I Uniform quantization with deadzone is used to quantize all the wavelet coefficients. For each subband b, a basic quantizer step size b is selected by the user and is used to quantize all the coefficients in that subband. The choice of the quantizer step size for each subband can be based on visual models and is likened to the q-table specification in the JPEG DCT. b b b 2 b b b b -3-2 -1 0 1 2 3 32
Embedded Quantization in Part I Unlike JPEG Baseline, where the resulting quantizer index q is encoded as a single symbol, in JPEG2000 it is encoded one bit at a time, starting from the MSB and proceeding to the LSB. During this progressive encoding, the quantized wavelet coefficient is called insignificant if the quantizer index q is still zero. Once the first nonzero bit is encoded, the coefficient becomes significant and its sign is encoded. If the p least significant bits of the quantizer index still remain to be encoded, the reconstructed sample at that stage is identical to the one obtained by using a UTQ with deadzone with a step size of b 2 p. 33
Embedded Quantization by Bit-Plane Coding 2 N-1 2 N-2 2 N-3.. 2 0 0 1 1 0 MSB LSB 34
Embedded Quantization Example Wavelet coefficient = 83 Quantizer step size = 3 Quantizer index = 83/3 = 27 Quantizer index in binary: 00011011 Decoded index after 6 BP s: 000110 = 6 Step size with 2 BP s remaining = 12 Quantizer index with step size of 12 = 6 Dequantized value = (6 + 0.5) x 12 = 78 35
JPEG2000 Part I Encoder Original Image Data Discrete Wavelet Transform (DWT) Uniform Quantizer with Deadzone Compressed Image Data Adaptive Binary Arithmetic Coder Context-based adaptive binary arithmetic coding is used in JPEG2000 to efficiently compress each individual bit plane. 36
JPEG2000 Entropy Coder Each bit plane is further broken down into blocks (e.g., 64 x 64). The blocks are coded independently (i.e., the bit stream for each block can be decoded independent of other data) using three coding passes. The coding progresses from the most significant bit-plane to the least significant bit-plane. A coding block of a bit plane of a subband 37
JPEG2000 Entropy Coder The binary value of a sample in a block of a bit plane of a subband is coded as a binary symbol with the JBIG2 MQ- Coder that is a context-based adaptive arithmetic coder. Each bit-plane of each block of a subband is encoded in three sub bit plane passes instead of a single pass. The bitstream can be truncated at the end of each pass. This allows for: Optimal embedding, so that the information that results in the most reduction in distortion for the least increase in file size is encoded first. A larger number of bit-stream truncation points to achieve finer SNR scalability. 38
Significance Propagation Pass The first pass in a new bit plane is called the significance propagation pass. A symbol is encoded if it is insignificant but at least one of its eight-connected neighbors is significant as determined from the previous bit plane and the current bit plane based on coded information up to that point. These locations have the highest probability of becoming significant). The probability of the binary value at a given location of a bit-plane of a block of a subband is modeled by a context formed from the significance values of its neighbors. d v d h X h d v d 39
Refinement and Clean-up Passes Refinement (REF): Next, the significant coefficients are refined by their bit representation in the current bit-plane. Clean-up: Finally, all the remaining coefficients in the bitplane are encoded. (Note: the first pass of the MSB bit-plane of a subband is always a clean-up pass). The coding for the first and third passes are identical, except for the run coding that is employed in the third pass. The maximum number of contexts used in any pass is no more than nine, thus allowing for extremely rapid probability adaptation that decreases the cost of independently coded segments. 40
Sig. Prop. = 0 Bit plane 1 Refine = 0 Cleanup = 21 Total Bytes 21 Systems Engineering Compression ratio = 12483 : 1 RMSE = 39.69 PSNR = 16.16 db % refined = 0 % insig. = 99.99 41
Sig. Prop. = 38 Refine = 13 Cleanup = 57 Total Bytes 108 Systems Engineering Bit plane 3 Compression ratio = 1533 : 1 RMSE = 21.59 PSNR = 21.45 db % refined = 0.05 % insig. = 99.89 42
Sig. Prop. = 224 Refine = 73 Cleanup = 383 Total Bytes 680 Systems Engineering Bit plane 5 Compression ratio = 233 : 1 RMSE = 12.11 PSNR = 26.47 db % refined = 0.23 % insig. = 99.43 43
Sig. Prop. = 1243 Refine = 418 Cleanup = 1349 Total Bytes 3010 Systems Engineering Bit plane 7 Compression ratio = 47 : 1 RMSE = 6.02 PSNR = 32.54 db % refined = 1.32 % insig. = 97.09 44
Sig. Prop. = 4593 Refine = 1925 Cleanup = 5465 Total Bytes 11983 Systems Engineering Bit plane 9 Compression ratio = 11.2 : 1 RMSE = 2.90 PSNR = 38.87 db % refined = 6.01 % insig. = 87.66 45
Sig. Prop. = 25421 Refine = 8808 Cleanup = 5438 Total Bytes 39667 Systems Engineering Bit plane 11 Compression ratio = 2.90 : 1 RMSE = 0.90 PSNR = 49.00 db % refined = 28.12 % insig. = 46.80 46
Wavelet Bit Plane Compression Table BP JPEG-2K Bytes JPEG-2K PSNR-dB JPEG-2K RMSE JPEG Bytes JPEG PSNR JPEG RMSE 1 21 16.16 39.69 2 63 18.85 29.11 3 171 21.45 21.59 4 442 23.74 16.58 5 1122 26.47 12.11 6 2601 29.39 8.65 7 5821 * 32.54 6.02 5804 * 29.77 8.28 8 11680 * 35.70 4.18 11696 * 33.33 5.49 9 23716 * 38.87 2.90 23932 * 36.50 3.81 10 51164 * 43.12 1.78 51636 * 40.16 2.50 11 90855 * 49.00 0.90 91216 * 44.12 1.59 6-level decomposition with (9,7) filter IJG code with default q-table * Filesize includes the header 47
Comparison of JPEG and JPEG-2000 JPEG, PSNR = 29.77 db Filesize = 5804 Bytes JPEG-2K, PSNR = 32.54 db Filesize = 5821 Bytes (BP 7) 48
Benefits of the JPEG2000 Entropy Coder Bit plane representation of data allows for embedded bit stream Region of interest coding (ROI) coding can be performed by prioritizing the coding of the ROI bit plane information Arithmetic coding allows for efficient compression of sparse binary data Context modeling allows for efficient compression of binary correlated data Packetized information allows for improved error resilience 49
Example of Bit Plane Reordering Subband 1 Subband 2 Subband 3 Subband 4 50
Example of Bit-Plane Reordering Code blocks BP 1 1 2 3 4 BP 2 5 9 13 6 10 14 7 11 15 8 12 16 41 42 43 44 69 70 71 72 BP 3 17 21 25 18 22 26 19 23 27 20 24 28 45 49 53 46 50 54 47 51 55 48 52 56 73 77 81 74 78 82 75 79 83 76 80 84 Significance Refinement Clean-up BP 4 29 33 37 30 34 38 LL 31 35 39 32 36 40 57 61 65 58 62 66 59 63 67 HL 60 64 68 85 89 93 86 90 94 LH 87 91 95 88 92 96 97 98 99 100 HH 51
Example of Bit-Plane Data Ordering MSB BP 1 BP 2 Code blocks BP 3 BP 4 Significance Refinement Clean-up BP 5 BP 6 LL2 HL2 LH2 HH2 HL1 LH1 HH1 52
Lowest Resolution, Highest Quality BP 1 BP 2 BP 3 BP 4 BP 5 BP 6 LL2 HL2 LH2 HH2 HL1 LH1 HH1 53
Medium Resolution, Highest Quality BP 1 BP 2 BP 3 BP 4 BP 5 BP 6 LL2 HL2 LH2 HH2 HL1 LH1 HH1 54
Highest Resolution, Highest Quality BP 1 BP 2 BP 3 BP 4 BP 5 BP 6 LL2 HL2 LH2 HH2 HL1 LH1 HH1 55
Highest Resolution, Target Visual Quality BP 1 BP 2 BP 3 BP 4 BP 5 BP 6 LL2 HL2 LH2 HH2 HL1 LH1 HH1 56
Resolution & Decomposition Levels Resolution level 0 Decomposition level 4 Decomposition level 3 4LL 4HL 4LH 4HH 3LH 3HL 3HH 2HL 1HL Decomposition level 2 2LH 2HH Decomposition level 1 1LH 1HH Decomposition level A set of HL, LH, and HH subbands. For the last level of decomposition (N L ), the LL subband is included. Decomposition levels run from 1 to N L. Resolution level Related to decomposition angle by n = N L r, where n is a decomposition level and r is a resolution level. Reconstruction levels run from N L to 0. 57
Code Blocks, Packets & Precincts Code blocks divide subbands into regions that may be extracted independently A rectangular grouping of coefficients from the same subband Code blocks do not scale with levels Packets Part of the compressed bitstream. Contains packet header & compressed image data from one layer of one precinct of one decomposition level Contains compressed image data from all code blocks of a given resolution level (nhl, nlh, and nhh subbands) Precinct Rectangular region within a decomposition level used to limit the size of packets Subbands with more than one precinct, have multiple packets in a given layer. Code block Subband boundary Precinct boundary 58
Layer Formation Least important bits Layer 5 4 3 2 Most important bits 1 1 2 4 Code block 3 5 6 7 Layer An arbitrary collection of compressed image data from coding passes of one or more code blocks. Typically a layer represents an improvement in image quality. Packet headers describe the contribution of each code block in each layer 59
JPEG 2000 Progression Types After wavelet processing, we have a four dimensional cube of data Spatial/Resolution (two) Component Bitdepth JPEG 2000 allows progression along four dimensions Layer (L) Resolution (R) Component (C) Precinct or position (P) These are roughly equivalent as follows Resolution & Precinct Spatial/Resolution Component Component Layer Bitdepth Spatial/Resolution Component Spatial/Resolution Bitdepth (4 th dimension) Wavelet processed components 60
JPEG 2000 Progression Types There are five allowed progression types LRCP, RLCP, RPCL, PCRL, CPRL View progression as four nested loops, read left to right LRCP Progression by SNR. Best full size image (SNR) Loop over all layers RLCP Loop over all resolution levels Loop over all components Loop over all precinct (positions) Progression by resolution. Improving image quality for a fixed resolution Note that C (component) is in an inner loop in all progressions except CPRL 61
Effects of Layering 55 50 45 One Layer Multiple Layers One Layer (lossless) Multiple Layer (lossless) One Layer vs. Multiple Layers (LRCP) PSNR (db) 40 35 30 25 20 0 0.5 1 1.5 2 2.5 3 3.5 Rate (bpppb) 62
Effects of Progression 55 50 45 RLCP LRCP RLCP Lossless LRCP Lossless RLCP vs. LRCP Comparison PSNR (db) 40 35 30 25 20 0 0.5 1 1.5 2 2.5 3 3.5 Rate (bpppb) 63
Tiles Tiles are independently coded sub images. Nothing crosses tile boundaries Wavelet Entropy coding Layers Progressions Tiles may be broken into tile parts. Tile parts from different tiles can be interspersed in a codestream Only mechanism available to achieve tile progression In general, need to parse data out of tiles to achieve a different image quality If all tiles are compressed at 2.0 bpp and you want 1.0 bpp, then need to go into each tile and get the 1.0 bpp 64
Bitstream Ordering with Four Quality Layers BP 1 BP 2 BP 3 BP 4 BP 5 BP 6 Layer 1 Layer 2 Layer 3 Layer 4 LL2 HL2 LH2 HH2 HL1 LH1 HH1 65
JPEG2000 Bitstream Syntax Precinct: Each subband is divided into non-overlapping rectangles of equal size (except maybe at the boundaries where the size can be different) called precincts. Precincts provide some level of spatial locality in the bit stream and their boundaries are aligned with code blocks. Their size can vary for each tile, (color) component, and resolution. Packet: Consists of the compressed bit streams associated with a certain number of sub-bit planes from each codeblock in a precinct. Packets serve as one quality increment for one resolution level at one spatial location. Layer: is a collection of packets, one from each precinct at each resolution. It can be interpreted as one quality increment for the entire image at full resolution. 66
Example: Precincts and Codeblocks Code blocks Precincts 67
JPEG 2000 Codestream Structure Delimiting markers These are markers that are used in the start of most major sections of codestream and the very end. Start of codestream (SOC), start of tile part (SOT), start of data (SOD), and end of codestream (EOC) Fixed information marker segments This marker includes information that is required to properly decode the image. Size Marker (SIZ) which includes reference grid size, tile size, resolution/sampling (relative to grid), image and tile offsets (into the grid), number of components, and component precision (data type and bit depth) 68
JPEG 2000 Codestream Structure Functional Marker Segments These marker define the parameters used in the compression of a tile or an image. The order or precedence of the markers are Tile Part Component Marker Tile Part Default Marker Image Component Marker Image Default Marker Default markers (can be used in image header or tile header) coding style default (COD), quantization default (QCD), region of interest (RGN), progression order changes (POC) Component Markers coding style component(coc), quantization component (QCC) 69
JPEG 2000 Codestream Structure Functional Markers Coding style default and coding style component include information on coding style, number of decompositions, code-block size, code-block style, wavelet transformation, precinct style Quantization default and quantization component include information on the quantization for the derived, expounded or no quantization Region of interest marker includes the location of the ROI Progression order changes describes the bounds and progression order for any progression other than that in the COD marker in the main image header 70
JPEG 2000 Codestream Structure Pointer Marker Segments Pointer markers are used for quick access to data that is required for decompression of a given location, resolution quality, or component. All of the pointer segments define lengths of segments which allow fast rearranging of data and pointer markers Main Header Markers Tile part lengths (TLM), packet lengths main header (PLM), packed packet headers (PPM) (main header). Tile part header Packet length tile-part header (PLT), packed packet headers (tile part header) (PPT). 71
JPEG 2000 Codestream Structure In Bit stream markers Start of packet (SOP) and end of packet (EOP) markers are used to isolate a given packet in a noisy environment Information marker segments Component registration (CRG) marker is used to register components if the components do not have the same sampling to the reference grid The comment marker (COM) is an open style marker that allows for unstructured data 72
JPEG 2000 Codestream Structure The J2K codestream starts with the main header, followed by tile-part header(s), and bitstream(s) and ends with an EOC marker The main header starts with SOC and SIZ markers, then followed (in any order) by COD and QCD markers and possibly QCC, RGN, POC, PPM, TLM, PLM, CRG, COM The tile part header starts with SOT marker and finishes with SOD and can contain (in any order) COD, COC, QCD, QCC, RGN, POC, PPT, PLT, COM The bitstream may contain SOP and EPH markers 73
Resolution Progressive Example All images have been decompressed from the same bit stream. The wavelet decomposition provides a natural resolution hierarchy. 74
SNR Scalability Example Original 8-bit image 8-to-1 Compression 16-to-1 Compression 32-to-1 Compression 64-to-1 Compression 128-to-1 Compression All images have been decompressed from the same bit-stream 75
Spatial Progression Example Image after decoding 3 precincts using the component-precinctresolution-layer ordering. Precinct sizes are chosen such that each subband precinct will correspond to a 1024x1024 image region. The image is (1024x2560), so there are 2 partial precincts at the bottom of the image. 76
Region of Interest (ROI) Coding Region-of-interest (ROI) coding allows selected parts of an image to be coded with higher quality compared to the background (BG). The ROI encoding is done using an ROI mask. This binary mask is generated in the wavelet domain and describes which quantized wavelet coefficients must be encoded with higher quality. It depends on: ROI region specification in the image domain DWT filter 77
ROI Coding in Part II Part II allows for the scaling method of ROI coding, where the bits representing the wavelet coefficients contributing to the ROI are shifted upward by a user-defined value. This Allows for coding the ROI with any desired quality differential compared to the background. Multiple ROI s are allowed, each with its own corresponding scaling value. The ROI shape is limited to rectangles and circles. ROI coordinates and shift values are signaled in the bitstream. The ROI mask needs to be generated both on the encoder and the decoder side. 78
Region of Interest (ROI) Example ROI has bit rate of 2.0 bpp Rest of image has bit rate of 0.0625 bpp Bit rate for entire image is 0.12 bpp 79
ROI Coding in Part I: Maxshift Method In the Maxshift method, the wavelet coefficients in the ROI region are scaled up by a fixed number of bits s, so that the smallest shifted nonzero ROI coefficient is larger than the largest BG coefficient. The parameter s is signaled in the bit stream. As a result, the decoder can discriminate between the ROI and BG coefficients by comparing each decoded value to a threshold. Pros and cons are: It allows for multiple regions of arbitrary shape ROI without the inclusion of the shape information in the bit stream or need for ROI mask generation at the decoder. The user can prioritize the coding of the ROI region over the BG but does not have control over the quality differential between ROI and BG. 80
Maxshift Method of ROI Coding Coefficient Value ROI Coefficients 81
Other JPEG2000 Features Tiling of large images provides for independent processing of different image regions. The canvas coordinate system allows for efficient recompression of cropped images. Rich bit stream syntax provides means for transcoding of the data for streaming, resolution progression, quality progression, or any mixture thereof. Rich file format allows for various color spaces and metadata information. Compressed domain image manipulation (rotations of 90, 180, 270 degrees, horizontal and vertical flipping) 82
JPEG2000 Feature Summary Improved coding efficiency (up to 30% compared to DCT) Multi-resolution representation Quality scalability (SNR or visual) Target bit rate (constant bit rate applications) Lossless to lossy progression Improved error resilience Tiling Rich bit stream syntax (layering, packet partitions, canvas coordinate system, etc.) Rich file format 83